• The Journal of Korea Robotics Society
• /
• v.17 no.2
• /
• pp.133-141
• /
• 2022

This paper introduces model compression algorithms which make a deep neural network smaller and faster for embedded systems. The model compression algorithms can be largely categorized into pruning, quantization and knowledge distillation. In this study, gradual pruning, quantization aware training, and knowledge distillation which learns the activation boundary in the hidden layer of the teacher neural network are integrated. As a large deep neural network is compressed and accelerated by these algorithms, embedded computing boards can run the deep neural network much faster with less memory usage while preserving the reasonable accuracy. To evaluate the performance of the compressed neural networks, we evaluate the size, latency and accuracy of the deep neural network, DenseNet201, for image classification with CIFAR-10 dataset on the NVIDIA Jetson Xavier.

• Proceedings of the Korean Geotechical Society Conference
• /
• 2006.03a
• /
• pp.1159-1164
• /
• 2006

In this paper an artificial neural network model is developed to estimate the unconfined compression strength of Deep Cement Mixing(DCM) treated soil. A database which consists of a number of unconfined compression test result compiled from 9 clay sites is used to train and test of the artificial neural network model. Developed neural network model requires water content of soil, unit weight of soil, passing percent of #200 sieve, weight of cement, w-c ratio as input variables. It is found that the developed artificial neural network model can predict more precise and reliable unconfined compression strength than the conventional empirical models.

• Journal of the Semiconductor & Display Technology
• /
• v.22 no.1
• /
• pp.129-133
• /
• 2023

Image compression is a fundamental technique in the field of digital image processing, which will help to decrease the storage space and to transmit the files efficiently. Recently many deep learning techniques have been proposed to promise results on image compression field. Since many image compression techniques have artifact problems, this paper has compared two deep learning approaches to verify their performance experimentally to solve the problems. One of the approaches is a deep autoencoder technique, and another is a deep convolutional neural network (CNN). For those results in the performance of peak signal-to-noise and root mean square error, this paper shows that deep autoencoder method has more advantages than deep CNN approach.

• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2022.11a
• /
• pp.112-115
• /
• 2022

The process of deep learning usually needs to deal with massive data which has greatly limited the development of deep learning technologies today. Convolutional Neural Network (CNN) structure is often used to solve image classification problems. However, a large number of images may be required in order to train an image in CNN, which is a heavy burden for existing computer systems to handle. If the image data can be compressed under the premise that the computer hardware system remains unchanged, it is possible to train more datasets in deep learning. However, image compression usually adopts the form of lossy compression, which will lose part of the image information. If the lost information is key information, it may affect learning performance. In this paper, we will analyze the effect of image compression on deep learning performance on CNN-based cat and dog classification. Through the experiment results, we conclude that the compression of images does not have a significant impact on the accuracy of deep learning.

• Journal of Information Processing Systems
• /
• v.17 no.2
• /
• pp.411-425
• /
• 2021

In this paper, a vehicle recognition algorithm based on deep convolutional neural network and compression dictionary is proposed. Firstly, the network structure of fine vehicle recognition based on convolutional neural network is introduced. Then, a vehicle recognition system based on multi-scale pyramid convolutional neural network is constructed. The contribution of different networks to the recognition results is adjusted by the adaptive fusion method that adjusts the network according to the recognition accuracy of a single network. The proportion of output in the network output of the entire multiscale network. Then, the compressed dictionary learning and the data dimension reduction are carried out using the effective block structure method combined with very sparse random projection matrix, which solves the computational complexity caused by high-dimensional features and shortens the dictionary learning time. Finally, the sparse representation classification method is used to realize vehicle type recognition. The experimental results show that the detection effect of the proposed algorithm is stable in sunny, cloudy and rainy weather, and it has strong adaptability to typical application scenarios such as occlusion and blurring, with an average recognition rate of more than 95%.

• Journal of the Institute of Convergence Signal Processing
• /
• v.22 no.3
• /
• pp.99-103
• /
• 2021

Recently, convolutional neural networks (CNNs) have been widely utilized due to their outstanding performance in various computer vision fields. However, due to their computational-intensive and high memory requirements, it is difficult to deploy CNNs on hardware platforms that have limited resources, such as mobile devices and IoT devices. To address these limitations, a neural network compression research is underway to reduce the size of neural networks while maintaining their performance. This paper proposes a CNN compression technique that dynamically adjusts the thresholds of pruning, one of the neural network compression techniques. Unlike the conventional pruning that experimentally or heuristically sets the thresholds that determine the weights to be pruned, the proposed technique can dynamically find the optimal thresholds that prevent accuracy degradation and output the light-weight neural network in less time. To validate the performance of the proposed technique, the LeNet was trained using the MNIST dataset and the light-weight LeNet could be automatically obtained 1.3 to 3 times faster without loss of accuracy.

• Smart Structures and Systems
• /
• v.32 no.3
• /
• pp.179-193
• /
• 2023

This study explores an alternative to the existing centralized process for data anomaly detection in modern Internet of Things (IoT)-based structural health monitoring (SHM) systems. An edge intelligence framework is proposed for the early detection and classification of various data anomalies facilitating quality enhancement of acquired data before transmitting to a central system. State-of-the-art deep neural network pruning techniques are investigated and compared aiming to significantly reduce the network size so that it can run efficiently on resource-constrained edge devices such as wireless smart sensors. Further, depthwise separable convolution (DSC) is invoked, the integration of which with advanced structural pruning methods exhibited superior compression capability. Last but not least, quantization-aware training (QAT) is adopted for faster processing and lower memory and power consumption. The proposed edge intelligence framework will eventually lead to reduced network overload and latency. This will enable intelligent self-adaptation strategies to be employed to timely deal with a faulty sensor, minimizing the wasteful use of power, memory, and other resources in wireless smart sensors, increasing efficiency, and reducing maintenance costs for modern smart SHM systems. This study presents a theoretical foundation for the proposed framework, the validation of which through actual field trials is a scope for future work.

• Journal of Broadcast Engineering
• /
• v.23 no.3
• /
• pp.383-394
• /
• 2018

In this paper, we investigate image and video compression techniques based on deep learning which are actively studied recently. The deep learning based image compression technique inputs an image to be compressed in the deep neural network and extracts the latent vector recurrently or all at once and encodes it. In order to increase the image compression efficiency, the neural network is learned so that the encoded latent vector can be expressed with fewer bits while the quality of the reconstructed image is enhanced. These techniques can produce images of superior quality, especially at low bit rates compared to conventional image compression techniques. On the other hand, deep learning based video compression technology takes an approach to improve performance of the coding tools employed for existing video codecs rather than directly input and process the video to be compressed. The deep neural network technologies introduced in this paper replace the in-loop filter of the latest video codec or are used as an additional post-processing filter to improve the compression efficiency by improving the quality of the reconstructed image. Likewise, deep neural network techniques applied to intra prediction and encoding are used together with the existing intra prediction tool to improve the compression efficiency by increasing the prediction accuracy or adding a new intra coding process.

• International Journal of Computer Science & Network Security
• /
• v.23 no.11
• /
• pp.73-76
• /
• 2023

We introduce DCNN and DRAE appraoches for compression of medical videos, in order to decrease file size and storage requirements, there is an increasing need for medical video compression nowadays. Using a lossy compression technique, a higher compression ratio can be attained, but information will be lost and possible diagnostic mistakes may follow. The requirement to store medical video in lossless format results from this. The aim of utilizing a lossless compression tool is to maximize compression because the traditional lossless compression technique yields a poor compression ratio. The temporal and spatial redundancy seen in video sequences can be successfully utilized by the proposed DCNN and DRAE encoding. This paper describes the lossless encoding mode and shows how a compression ratio greater than 2 (2:1) can be achieved.

• The Journal of the Acoustical Society of Korea
• /
• v.37 no.6
• /
• pp.475-482
• /
• 2018

In this paper, we present a study on data augmentation methods for DNN (Deep Neural Network)-based audio tagging. In this system, an audio signal is converted into a mel-spectrogram and used as an input to the DNN for audio tagging. To cope with the problem associated with a small number of training data, we augment the training samples using time stretching, pitch shifting, dynamic range compression, and block mixing. In this paper, we derive optimal parameters and combinations for the augmentation methods through audio tagging simulations.